agentic-genai-maintenance-copilot

Agentic GenAI Maintenance Copilot is an Azure-deployed Agentic GenAI system for grounded rotating-equipment maintenance investigation, evidence retrieval, diagnosis, and human-approved work-order actions. It simulates an industrial maintenance copilot that investigates equipment issues using multiple evidence sources instead of relying on an LLM alone. The agent can query asset data, analyze sensor trends, retrieve maintenance history and engineering documents, produce a citation-backed diagnosis, and propose a controlled work order when the evidence justifies one.

State-changing application actions are protected by a human-in-the-loop approval workflow. The system can pause a LangGraph investigation, wait for an operator decision, validate the approval, resume from the saved checkpoint, and complete the workflow without claiming that physical maintenance was performed. The copilot does not directly control machinery, shut down equipment, modify PLC parameters, or bypass equipment interlocks. All V3 investigation tools are read-only.

The implementation includes a LangGraph model-tool-model workflow with bounded iterations, Azure OpenAI gpt-5.4-mini with structured outputs, Sentence Transformers embeddings and Chroma engineering-document retrieval, a FastAPI REST API, and a Streamlit operator dashboard. It also includes a 15-scenario deterministic evaluation suite, 529 automated tests, Ruff, GitHub Actions CI, and deployment with Docker, Azure Container Registry, Azure Container Apps, and Bicep.